A Study on the Development of Deepfake-Based Deep Learning Algorithm for the Detection of Medical Data Manipulation

Abstract

In the medical field, the abuse of manipulation data through image processing technology of deep learning is fatal. Therefore, research on detection of modulation on medical images is essential. The data set for fundus data manipulation used 356 right fundus images of 4 lesions (normal, diabetic retinopathy, glaucoma, macular degeneration) out of about 6,000 data collected by Shangong Medical Technology Co., Ltd. The training and verification dataset of the manipulation detection model used original data and U-Net manipulation data. In addition, data manipulated in the Cycle General Adversarial Network (GAN) model were used for the diversity of verification. In this paper, three ophthalmologists and two general doctors were asked to verify the above modulation data. Verification was requested for each lesion, and the verification results were shown through the Receiver operating characteristic (ROC) curve and the Area Under the Curve (AUC). The verification of this study evaluated a total of 100 randomly extracted manipulation data and original data as Observer Performance Test (OPT) for each group. When the evaluation results were digitized as average scores, the scores of ophthalmologists group: 0.72 and general doctors group: 0.67 were recorded. The manipulated images were so similar that both ophthalmologists and general doctors could not find about 30%. However, the manipulation detection model studied in this paper was excellent in about 20% of the group OPT score with a lesion average of 0.913 in the same data group. Therefore, it can be seen that the manipulation detection model of this study finds the manipulated image and the original image well. The future plan is to expand the scope of manipulation detection data to conduct research on various medical data. After that, it will verify its availability at the actual site.